EI、Scopus 收录
中文核心期刊
方培俊, 蔡英凤, 陈龙, 孙晓强, 王海. 基于ED-LSTM的智能汽车神经网络横向动力学建模与控制. 力学学报, 2022, 54(7): 1896-1908. DOI: 10.6052/0459-1879-21-667
引用本文: 方培俊, 蔡英凤, 陈龙, 孙晓强, 王海. 基于ED-LSTM的智能汽车神经网络横向动力学建模与控制. 力学学报, 2022, 54(7): 1896-1908. DOI: 10.6052/0459-1879-21-667
Fang Peijun, Cai Yingfeng, Chen Long, Sun Xiaoqiang, Wang Hai. Neural network lateral dynamics modeling and control based on ED-LSTM for intelligent vehicle. Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(7): 1896-1908. DOI: 10.6052/0459-1879-21-667
Citation: Fang Peijun, Cai Yingfeng, Chen Long, Sun Xiaoqiang, Wang Hai. Neural network lateral dynamics modeling and control based on ED-LSTM for intelligent vehicle. Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(7): 1896-1908. DOI: 10.6052/0459-1879-21-667

基于ED-LSTM的智能汽车神经网络横向动力学建模与控制

NEURAL NETWORK LATERAL DYNAMICS MODELING AND CONTROL BASED ON ED-LSTM FOR INTELLIGENT VEHICLE

  • 摘要: 车辆动力学建模过程中通常会进行简化和假设, 导致模型在某些工况下无法准确反映车辆的实际动态特性, 影响控制精度甚至安全性. 鉴于此, 该文提出了一种基于数据驱动的非线性建模与控制方法, 建立了新型神经网络车辆横向动力学多步预测模型, 实现了智能汽车对参考轨迹的跟踪控制. 首先, 在分析车辆单轨模型并考虑轮胎非线性和纵向负载转移的基础上, 基于编码器−解码器结构设计神经网络横向动力学模型. 其中, 使用串行排列来扩展微分方程描述不完全的动力学信息, 隐藏层神经元学习车辆的高度非线性和强耦合特性, 进而提高模型全局计算精度. 利用所构建的数据集进行模型训练和测试, 结果表明, 相比于物理模型, 所提出的模型在不同路面附着系数条件下均具有更高的建模精度, 具有隐式预测路面摩擦条件能力. 其次, 利用提出的模型设计轨迹跟踪控制算法, 根据车辆稳态转向假设, 计算所需的前轮转向角和稳态质心侧偏角, 将稳态质心侧偏角纳入基于路径误差的转向反馈中, 实现参考轨迹跟踪控制. 最后, 使用CarSim/Simulink联合仿真及HIL实验测试进行不同工况试验的对比分析, 对所提出的基于神经网络模型的控制算法进行评价, 结果表明, 该模型能够实现智能汽车在高速下精确的跟踪控制效果, 并具有良好的横向稳定性.

     

    Abstract: Simplification and assumptions are usually made in the process of vehicle dynamics modeling, resulting in the model can not accurately reflect the actual dynamic characteristics of the vehicle under some working conditions, which affects the control accuracy and even safety. In view of this, this paper proposes a data-driven nonlinear modeling and control method, establishes a new neural network multi-step prediction model of vehicle lateral dynamics, and realizes the tracking control of reference trajectory of intelligent vehicle. Firstly, based on the analysis of vehicle single-track model and considering tire nonlinearity and longitudinal load transfer, a neural network lateral dynamics model is designed based on encoder decoder structure. Among them, the serial arrangement is used to expand the differential equation to describe the incomplete dynamic information, and the hidden layer neurons learn the highly nonlinear and strong coupling characteristics of the vehicle, so as to improve the global calculation accuracy of the model. Use the constructed data set for model training and testing, the results show that, compared with the physical model, the proposed model has higher modeling accuracy under different road adhesion coefficients, and has the ability to implicitly predict road friction conditions. Secondly, using the proposed model to design trajectory tracking control algorithm, according to the vehicle steady-state steering assumption, the required front wheel steering angle and steady-state sideslip angle are calculated, and the steady-state sideslip angle is incorporated into the steering feedback based on path error to realize the reference trajectory tracking control. Finally, comparative analysis of tests under different working conditions is carried out with Simulink/CarSim co-simulation and HiL experiments to evaluate the proposed control algorithm based on neural network model. The results show that the model can realize the accurate tracking control effect of intelligent vehicle at high speed, and has good lateral stability.

     

/

返回文章
返回